Executive Summary : | The computing strength of biological neural networks is derived from analog processing and the potential of silicon technology to be better leveraged by utilizing device physics for computation. Machine learning (ML) approaches have significantly improved automated data processing and pattern identification in various sectors, including acoustics. This proposal is motivated by the potential influence of machine learning-based approaches in acoustics. As battery size impacts system power and energy usage, it is crucial to minimize power used for acoustic signal processing as the system gets smaller. A voice activity detector (VAD) is often used as a system wake-up mechanism to run power-hungry processes after the VAD wakes up the system. Most VADs consist of a feature extractor and classifier, but reducing VAD power is challenging due to computationally expensive feature extraction. Ultra-low-power acoustic classifiers will employ an in-filter computer architecture to extract features, using energy produced at the end of a filter's run over a certain time frame as a signal characteristic. Biasing MOS transistors at the subthreshold region will result in high energy efficiency. To implement the acoustic classifier, open-source EDA tools environment will be utilized. This approach can create a significant academic gap between those who can afford software and those who cannot. Open-source flows like qflow and companies like efabless will be employed to tapeout chips. The vision of open-source EDA tools in RTL-to-GDS using SKY130 PDK and the Semicon India initiative has inspired the PI and Co-PI to work on this topic. |